Generally, a knock decision level V.sub.KD is produced by multiplying a mean level of a knock sensor signal by a constant K. In such a KCS (knock control system), there is a problem that an optimum value of K changes due to manufacturing disperison, variations with time, etc. of engines, knock sensors and so on, which makes it difficult to perform accurate knock detection. In order to solve this problem, a system has been proposed which pays attention to the fact that, as shown in FIG. 15, the shape of a frequency distribution curve (broken line) of representative values of a knock sensor signal, in the case where a knock is generated, is distorted in the direction of a region of greater representative values of the knock sensor signal, in the case where a knock is generated, is distorted in the direction of a region of greater representative values of the knock sensor signal as compared with that of a frequency distribution curve (solid line) in the case where no knock is generated, and which contemplates to automatically correct a knock decision level on the basis of a degree of the distortion (for example, a relation in magnitude between the area of a region 1 and that of a region 2 in FIG. 15) (for example, see JP-A-62-267574).
However, when the region 1 or region 4 in FIG. 15 is used as a parameter for correcting the knock decision level as in the above system, a long time is required before a sufficient amount of data for correcting the knock decision level is obtained, since the probability for a representative value to fall within the region 1 or region 4 is small. Accordingly, there is a problem that a long time would elapse before the knock decision level is automatically corrected to have an adequate level and, as a result, inadequate knock detection is made during that time.